This reading group surveys some key papers.
Caretaker: ?? (hopefully some RF can volunteer here)
Monday, June 3: https://arxiv.org/abs/1902.04742
. A brief summary of discussion: for the Shamir-Yehudai paper, one interesting point is that there was doubt about whether a different treatment of the biases would lead to a different hardness result for the single relu learning (still a separation, but now merely polynomial. For the generalization paper, we all agreed that existing bounds are inadequate, but were not sure how to proceed, and it was unclear what the limits of uniform convergence are from here.